Heterogeneity (I²)

I² is derived from Cochran's Q, which compares the observed spread of study estimates with the spread expected from within-study sampling error alone. Because Q is influenced by the number and precision of studies, I² is not perfectly stable in very small meta-analyses; that is one reason expert guidance recommends interpreting it alongside the forest plot, confidence intervals, and the clinical similarity of the studies, not in isolation.

When the diamond looks tidy but the studies do not

One of the strangest moments in a meta-analysis is seeing a clean pooled estimate at the bottom of a forest plot while the individual studies above it are scattered all over the place. That is the trap Heterogeneity (I²) was built to expose. The bottom-line diamond can look calm even when the studies are not really telling the same story.

It is not “how big the effect is”

Here is the surprise: I² does not measure whether a treatment works a lot or a little. It measures how much the studies disagree with each other beyond what you would expect from chance alone.

Picture a choir trying to hold one note. If the voices cluster tightly, the average note represents the group well. If the voices drift apart, the average note still exists, but it stops sounding like any real singer. That is what high heterogeneity means in a meta-analysis: the pooled average may be mathematically correct while being a poor summary of what happened across studies.

So what does i2 represent? In plain language, it is the share of the total variation across study results that likely reflects real between-study differences rather than random sampling error.

What the numbers usually mean

I² is reported from 0% to 100%. A value near 0% means little observed inconsistency. Higher values mean more disagreement. The Cochrane Handbook gives rough, context-dependent guideposts: 0%-40% may not be important, 30%-60% may represent moderate heterogeneity, 50%-90% may represent substantial heterogeneity, and 75%-100% may represent considerable heterogeneity. These bands overlap on purpose because interpretation depends on the size and direction of effects, the populations studied, and how precise the studies are.

That is why there is no universal answer to “what is a good I2 heterogeneity?” Lower is easier to summarize, but “good” depends on context. In nutrition and supplement research, where trials often use different doses, durations, and participant groups, some heterogeneity is normal.

How do you calculate with I²?

The usual heterogeneity i² formula is based on Cochran’s Q: I² = 100% × (Q − df) / Q, where df means degrees of freedom, usually the number of studies minus one. If that math produces a negative number, it is set to 0% in practice. You do not usually calculate it by hand unless you are doing the meta-analysis yourself; most software reports it automatically.

What high heterogeneity should make you do next

If heterogeneity i² meaning is “the studies are not lining up neatly,” then the next move is not to throw the meta-analysis away. It is to ask why. Different doses, different baseline health status, different outcome definitions, short vs long trials, or higher risk of bias can all widen the spread.

One useful decision today: if you are reading a supplement meta-analysis and see I² above about 75%, do not stop at the pooled estimate. Look immediately for subgroup analyses or sensitivity analyses. If the paper cannot explain the spread, treat the average effect as a rough sketch, not a universal rule.